显示HN:Relay – 开源的Claude Cowork,适用于OpenClaw
Show HN: Relay – The open-source Claude Cowork for OpenClaw

原始链接: https://github.com/SeventeenLabs/relay

## Relay:您本地的自主AI指挥中心 Relay 是一款桌面应用程序,旨在赋予用户自主任务执行的能力——类似于 Claude Cowork,但拥有对数据、模型和治理的完全控制权。虽然 Claude Cowork 非常适合个人使用,但 Relay 解决了企业面临的关键限制:数据主权、模型锁定和合规性。 Relay 与 OpenClaw 运行时环境协同工作,允许代理在*您的*基础设施上运行(本地机器、VPS 或自定义 URL),使用*您*选择的 LLM(Claude、GPT-4、Llama、Gemini 或自定义端点)。它为操作员提供了一个控制平面,用于定义任务、审查交付成果、批准风险操作并跟踪成本。 **主要功能包括:** 调度、代理内存管理、连接器集成(Slack、Notion 等)以及完全可导出的审计跟踪以实现合规性。 Relay 提供精细的控制,对敏感操作(如文件删除或 shell 命令)设置审批门槛。 **适合:** 需要数据隐私、模型灵活性、可审计性和成本管理的企业。它是基于云解决方案的强大替代方案,让*您*掌控 AI 工作流程。

对不起。
相关文章

原文

Relay operator desk

What Is Relay · Why Relay · Features · Use Cases · Quickstart · Development

MIT License Electron React + Vite TypeScript

If OpenClaw is the runtime, Relay is your local command center.

Relay is an Electron desktop app that gives you the same experience as Claude Cowork — autonomous task execution, scheduling, sub-agents, connectors — but on your infrastructure, with your model, and with real governance.

Claude Cowork is excellent. But companies are looking for alternatives because of three structural limits:

  1. Data sovereignty — Cowork runs files through a sandboxed VM on Anthropic's servers. Regulated industries can't send sensitive data there.
  2. Model lock-in — Cowork only works with Claude. Companies want to route tasks to GPT-4, Llama, Gemini, or custom endpoints.
  3. Compliance gaps — Anthropic themselves recommend against using Cowork for regulated workflows because activities aren't yet captured in standard audit logs or compliance APIs.

Relay solves all three. Same workflow pattern — different trust model.


┌─────────────────────────────────────────────────────────────────────────┐
│                          You (Operator)                                  │
│         give goals · review deliverables · approve risky actions         │
└─────────────────────────────┬───────────────────────────────────────────┘
                              │
              ┌───────────────▼───────────────┐
              │     Relay (Desktop App)        │
              │     ── Control Plane ──        │
              │                                │
              │  Dispatch & Chat               │  You see everything.
              │  • Give a task in natural lang │  You approve what matters.
              │  • Agent plans steps for you   │  You stay in control.
              │  • Review polished deliverable │
              │                                │
              │  Governance                    │
              │  • Approval gates (file ops,   │
              │    shell commands, data sends) │
              │  • Exportable audit trail      │
              │  • Cost tracking per task      │
              │                                │
              │  Configure                     │
              │  • Schedule recurring tasks    │─ ─ ┐ Relay defines.
              │  • Browse & edit agent memory  │    │ OpenClaw executes.
              │  • Manage connectors (Slack,   │    │
              │    Notion, GitHub, Jira, etc.) │    │
              │  • Set project working folder  │    │
              └───────────────┬───────────────┘    │
                              │                    │
                       WebSocket / API             │
                              │                    │
              ┌───────────────▼───────────────┐    │
              │  OpenClaw Gateway (Runtime)    │◄ ─ ┘
              │  local · VPS · custom URL      │
              │  ── Execution Plane ──         │
              │                                │
              │  Agent Runtime                 │  Runs on YOUR infra.
              │  • Autonomous task execution   │  Your keys. Your data.
              │  • Multi-step planning & tools │
              │  • Sub-agent orchestration     │
              │                                │
              │  Persistence                   │
              │  • Memory storage & retrieval  │
              │  • Schedule runner (cron)      │
              │  • File read / write / search  │
              │                                │
              │  Integrations                  │
              │  • Connectors (Slack, Notion,  │
              │    GitHub, Jira, email, etc.)  │
              │  • Computer use (browser, UI)  │
              │  • Shell / script execution    │
              │                                │
              │  Model Router                  │
              │  • Routes to any LLM backend   │
              └──┬──────────┬──────────┬──────┘
                 │          │          │
           ┌─────▼──┐ ┌────▼───┐ ┌────▼─────┐
           │ Claude  │ │ GPT-4  │ │  Llama   │
           │ Gemini  │ │ Mixtral│ │  Custom  │
           └────────┘ └────────┘ └──────────┘

Relay is the control plane — you see, configure, and approve. OpenClaw is the execution plane — agents run, remember, and act on your infrastructure.

Example: Scheduled daily briefing

  Relay (you define)                 OpenClaw (it executes)
  ──────────────────                 ──────────────────────
  Create schedule: "Daily 8am"  ────►  Stores schedule
  Set connectors: Slack + Notion     Cron fires at 8am
                                     Agent reads project files
                                     Pulls Slack threads & Notion pages
                                     Calls LLM (your model choice)
                                     Writes briefing to memory
  Briefing appears in Relay  ◄────── Returns deliverable
  You review, approve, or redirect
  Full audit trail exported

Problem Relay's Answer
Data sovereignty Your files never leave your machine. Agents run on your server. Your keys.
Model lock-in Use any LLM through OpenClaw. Claude, GPT-4, Llama, custom endpoint — your choice.
No compliance-ready audit Full audit trail with exportable execution history, approval records, and action rationale.
Always-on execution Agents run 24/7 on a VPS while you control them from the desktop.
Token cost at scale Cowork burns through plan limits fast. Relay + OpenClaw lets you manage costs with your own infrastructure.
No syncing friction Your workspace files are in agent context in real-time. No FTP, no SSH, no copy-paste.

You give the goal → Agent plans the steps → You approve what matters → Agent executes → Everything is logged
Step What Happens
01 Connect Point Relay to your OpenClaw gateway (local, VPS, or custom). Verify health.
02 Dispatch Give the agent a task in a project context. Agent plans and starts working.
03 Approve High-risk actions pause for your review. You approve, reject, or redirect.
04 Track Full timeline: every action, approval, cost, and result — exportable.

  • ✅ You run OpenClaw on a VPS and want a desktop control plane
  • ✅ You need data sovereignty (GDPR, HIPAA, or internal policy)
  • ✅ You work in regulated industries (finance, legal, healthcare, government)
  • ✅ You want model-agnostic routing, not Claude-only lock-in
  • ✅ You need exportable audit trails for compliance
  • ✅ You want to control token costs on your own infrastructure

Not for you if:

  • ❌ You're happy with Claude-only on Anthropic's cloud (just use Cowork)
  • ❌ You want fully autonomous agents with zero human oversight

Why Not Just Use a Chat App?

You could wire OpenClaw to Telegram, Discord, or Slack and talk to your agent there. Many people do. Here's why that breaks down:

Capability Chat app (Telegram, etc.) Relay
Send a message to an agent ✅ Works ✅ Works
Approve risky actions before they run ❌ No approval gates — agent just does it ✅ File deletes, shell commands, data sends pause for review
See what the agent actually did ❌ You get a text reply, not an execution log ✅ Full timeline: every action, tool call, file change, cost
Schedule recurring tasks ❌ You'd need a separate cron + glue code ✅ Define schedules in the UI, OpenClaw runs them
Project context ❌ Every message starts from zero ✅ Tasks scoped to a working folder with persistent context
Agent memory ❌ Chat history is all you get ✅ Structured memory that persists across sessions
Manage connectors ❌ You wire each integration yourself ✅ Configure Slack, Notion, GitHub, Jira from the UI
Audit trail ❌ Scroll through chat logs ✅ Exportable execution history with approval records
Cost visibility ❌ No idea what a task cost ✅ Token usage and cost tracked per task
Multi-step execution ❌ Agent replies in one shot ✅ Agent plans steps, uses tools, reports back with deliverables

The bottom line: A chat app gives you a text box. Relay gives you an operator desk — dispatch, govern, track, and audit everything your agent does.


Native Electron app with persistent local state. Reliable day-to-day operations, not a browser tab. Dispatch tasks, guide decisions, and review results in one unified interface. Every task scoped to a working folder. No context drift between runs.
File deletes, shell commands, data exports — risky actions pause for your review. Every action logged with execution timeline, rationale, and approval records. Exportable. Connect to local, VPS, or custom OpenClaw-compatible endpoints. Any model.
Persistent operator context injected into every interaction. Agents remember. Create recurring tasks from the UI. Daily reports, weekly cleanups, continuous monitoring. Files, activity, memory, schedule, safety, and approvals in one operator shell.

Capability Relay Claude Cowork
Autonomous task execution
Scheduling
Sub-agents / multi-agent
Connectors / integrations ✅ (Anthropic cloud)
Desktop app
Local file access ✅ Truly local ⚠️ Sandboxed VM on Anthropic's servers
Self-hosted runtime
Model choice ✅ Any model via OpenClaw ❌ Claude only
Compliance-ready audit trail ✅ Exportable ❌ Not in audit logs or compliance APIs yet
Approval gates ✅ Per-action risk scopes ⚠️ Limited
Data on your infrastructure
Token cost control ✅ Your infra, your budget ⚠️ Plan limits, high token burn

The bottom line: Claude Cowork is excellent for personal productivity on Anthropic's cloud. Relay is for teams and companies that need data sovereignty, compliance-ready audit trails, and model freedom — all on their own infrastructure.

Note: Anthropic recommends against using Cowork for strongly regulated workflows because activities are not yet captured in standard audit logs or compliance APIs. Relay is built for exactly this gap.


Operations: Daily Briefing

Schedule a daily task → Agent synthesizes metrics, customer feedback, and team updates overnight → Results appear in Relay each morning → Operator reviews and acts on decision-ready recommendations.

Finance: Expense Approval

Agent flags an exception in an expense report → Relay pauses for approval → Finance lead reviews context and risk level → Approves or requests clarification → Action executes with full audit receipt.

Compliance: Recurring Audit Prep

"Every Friday, scan all project changes and produce a compliance summary" → Agent runs on schedule → Results appear in Relay → Exportable audit trail ready for review.

Technical: Code Review Automation

Agent runs npm test, scans for TODO comments, and produces a summary → Shell commands require approval → Operator reviews results → Follow-up actions triggered from one thread.

Product: Feedback Synthesis

Customer feedback scattered across email, Slack, and support tickets → Agent collects and synthesizes → Clusters themes with priority logic → Produces actionable roadmap recommendations.

"Every Friday, summarize trending topics in our space and draft 3 content ideas" → Results appear Monday morning → Team reviews and approves → Straight into the editorial calendar.


If you need... Use Cowork Use Relay
Personal AI productivity ✅ Great fit Overkill
Data sovereignty (GDPR, HIPAA) ❌ Data goes to Anthropic ✅ Your infrastructure
Compliance-ready audit logs ❌ Not available yet ✅ Built in, exportable
Model flexibility ❌ Claude only ✅ Any model via OpenClaw
Token cost control ❌ Plan limits ✅ Your infra, your budget
Team approval workflows ⚠️ Limited ✅ Per-action risk scopes
Always-on agents on your VPS ❌ Anthropic's cloud ✅ Your server, 24/7

Requirements:

Install & run:

git clone https://github.com/SeventeenLabs/relay.git
cd relay
npm install
npm run dev

This starts Vite, compiles Electron in watch mode, and launches Relay.

Optional cloud auth setup:

Set these when needed:

  • VITE_SUPABASE_URL
  • VITE_SUPABASE_ANON_KEY

In Settings → Gateway:

  1. Enter your OpenClaw gateway URL and token
  2. Save
  3. Run the health check

Typical endpoint patterns:

  • Local: ws://127.0.0.1:18789
  • VPS: wss://your-domain.com
  • Custom: Any OpenClaw-compatible endpoint

npm run dev                 # Full desktop dev loop
npm run build               # Build renderer + electron
npm run preview             # Preview renderer build
npm run package             # Build and package app to release/
npm run lint                # ESLint
npm run typecheck           # TS type checks (renderer + electron)
npm run verify              # lint + typecheck + smoke tests
npm run test:local-actions  # Local actions smoke tests
npm run test:e2e            # Electron E2E tests (mock gateway)


  • License: MIT
  • Copyright © 2026 SeventeenLabs
  • Built for operators who believe AI should be governed, auditable, and under human control.

Get started: Download the latest build or clone and run locally.

Questions? Open an issue or reach out to [email protected].

联系我们 contact @ memedata.com